The major weakness of mammography is the detection of cancer in women with radiographic dense breasts. While nearly all cancers will be apparent in fatty breasts, many fewer will be visible in extremely dense breast. This is due to masking of noncalcified cancers by surrounding dense tissue. Digital breast tomosynthesis (DBT) mammography is a new technology derived from digital mammography that was approved by the FDA in 2011 to improve detection and
characterization of breast lesions especially in women with nonfatty breasts (5). In DBT, the source x-ray tube is moved through a limited arc angle while the breast is compressed and a series of exposures are obtained (6). To a patient, DBT will be very similar to conventional digital mammography except there will be some movement of the x-ray tube head during exposures. These individual exposures are only a fraction of the total dose used during conventional digital mammography. The image data sets are reconstructed and the clinical reader is presented with a series of images (slices) through the entire breast that are read at a workstation analogous to CT or MRI study. Because each reconstructed slice may be as thin as 0.5 mm, masses and mass margins that may otherwise be superimposed with out-of-plane structures may be more visible in the reconstructed slice (Fig. 11-2). This should allow better visualization and characterization of noncalcified lesions. While the basic image interpretation will be similar to conventional mammography, new recall thresholds and probably benign thresholds will be established for DBT specific findings. In early studies, DBT has shown the ability to increase both sensitivity and specificity and has the potential to dramatically change not only how routine “mammography” is performed but also improve the clinical outcome of mammographic screening (5).

Interpretation of mammographic images involves the art and science of medicine. While the recognition and characterization of classic large tumors is often straightforward, the detection of the small, subtle lesions can challenge the most expert reader. Interpretive variability exists for screening and diagnostic mammography. Key factors that influence overall performance include physician expertise, recall rates, observation time, biopsy rates, double reading, and CAD. The relationships among these parameters are complex.

Similar to other areas of human endeavor and medicine, differences have been found among radiologists interpreting mammograms (8, 9, 10, 11, 12, 13 and 14). Beam et al., using an experimental model, found variation among practicing American radiologists with overall sensitivity ranging from 59% to 100% and specificity 35% to 98% (11). Sickles and colleagues reported higher cancer detection rates for specialists than generalists (6.0 per 1,000 vs. 3.4 per 1,000) within a single academic center in a retrospective clinical study (10). Specialists had higher volumes, more frequently participated in CME programs and fellowship training, and more often participated in radiologic-pathologic correlation conferences than generalists. The influence of reading volume on performance has not been consistent. Beam et al. tested 100 radiologists with an enriched study set of 148 mammograms with a 43% cancer incidence (11). They found reading volume not to be tightly associated with improved sensitivity. Rather complex multifactorial processes were found to be associated with expertise. Miglioretti and colleagues reported better performance for readers of diagnostic mammography at academic centers, those concentrating their time in breast imaging, and those performing breast biopsies (14). Volume was not
associated with performance. To date, no definite set of parameters completely predict reader expertise. Common associations with favorable interpretive skills and expertise include concentration in breast imaging, academic practice, continuing education, association with a multidisciplinary breast center, and practice audits (10, 12, 14). Reading a minimum of 480 mammographic cases per year is required by the FDA to maintain certification.

Sensitivity and specificity are inversely related for any particular reader due to nonspecific appearance of early breast cancer. Sensitivity increases with recall rate over a range of recall rates. High sensitivity can be achieved only when a sufficient number of women are recalled from screening for additional diagnostic mammography and ultrasonography (9, 13, 15, 17). The ideal balance between sensitivity and recall rate is controversial and reflects philosophy, cost, cultural issues, and medical-legal issues. Yankaskas and colleagues demonstrated sensitivity increased from 65% at recall rates of 1.9% to 4.4% and to 80% sensitivity at recall rates of 8.9% to 13.4% in a study of practicing North Carolina radiologists (15). Karssemeijer et al., in an enriched study population, found sensitivity for masses improved from approximately 35% at a 3% callback rate to 59% at a 20% callback rate (9). Gur and colleagues noted improvements in sensitivity with increasing recall over a wide range (7.7% to 17.2%, p < .05) at a large clinical practice (17). On average, a 0.22 per 1,000 cancer detection rate improvement occurred for every 1% absolute increase in recall rate. Otten et al., in an experimental situation, found 47% sensitivity improvement when FP rate increased from 1% to 4% (16). In the United States, callback rates of 5% to 15% are common. Rosenberg et al. reported the middle 50% recall rate for practicing U.S. radiologists to be 6.4% to 13.3% (mean 9.8%) for 2.5 million screening studies (13). European callback rates are frequently lower. Dutch breast cancer screening program has reported callback rates as low as 1.1% (9, 16). Emphasis on specificity and low cost will limit recall rate. Emphasis on high sensitivity will increase callback rates.

Mammographic sensitivity increases with reader observation time. Nodine and colleagues noted experienced mammographers made 71% of detections in the first 25 seconds but had continued true positive detections for approximately 80 seconds, albeit at a slower rate (18). Krupinski observed the detection of subtle findings occurred later in observation cycle than obvious masses which required longer visual dwell times (19). The threshold to initiate biopsy will influence cancer detection similar to recall rate thresholds. Higher thresholds for biopsy may be associated with higher false-negative (FN) rates and lower sensitivity (14).

Double reading (DR) has been advocated as a method to detect abnormalities overlooked by a single reader. Most independent DR studies have demonstrated improvement in sensitivity at a cost of lowered specificity. A review of clinical independent DR studies shows detection rate improvements of 4% to 15% (20). However, recall rates (FP) increased by 11% to 45%. These divergent trends for DR between sensitivity and specificity tend to balance accuracy. Taplin et al. showed independent DR improved sensitivity by 8.9% and decreased specificity by 13.6% similar to clinical trials (21). Accuracy, as determined by ROC methods, did not change suggesting that independent DR acted to shift the decision threshold towards sensitivity at the expense of specificity. DR with expert or consensus readers as the second reader appears to retain most improvements in sensitivity without large declines in specificity but this outcome may reflect in part the expertise of the second reader.

CAD systems, commercially available, use artificial computer intelligence in an attempt to act as a second reader. Like DR, most clinical CAD trials have shown improvement in sensitivity but declines in specificity. A CAD system functions as a second reader by placing “marks” on a mammographic site deemed suspicious. The radiologists then characterize these CAD detections. CAD can correctly identify approximately 60% to 80% of cancers with highest performance for microcalcifications. Unfortunately, CAD systems are very nonspecific with 2 to 4 marks placed per every exam. It has been estimated that only 1 in 5,000 CAD marks will reflect a true-positive finding representing a unique cancer missed by the radiologist (20). The interactions between radiologist and CAD are complex but tend to mirror DR studies. Clinical studies with CAD show sensitivity improvements varying from 1.7% to 19.5%, with declines in specificity as noted by increased recall rate of 0.1% to 26% (20). CAD effect on accuracy has often been incompletely reported so the determination whether CAD is increasing accuracy or shifting the threshold toward sensitivity has been questioned. Fenton et al. showed reader accuracy as measured by ROC methods declined with incremental use of CAD in a retrospective clinical study of 684,956 women (22). They showed a nonsignificant 6% improved sensitivity but a 13% significant loss of specificity. Sensitivity improvement for CAD was for DCIS. The reason for this negative result is uncertain but may result from overrelying on CAD, changing radiologists reading patterns, spending less overall time in observation, and radiologists being overwhelmed with the large number of false-positive CAD marks. While CAD remains controversial, CAD is in its infancy and the future of CAD is robust for detection and characterization tasks as a second reader. More effort will be required to improve the human interaction with CAD to reap the theoretical advantages of CAD. Current CAD systems are best viewed as a second reader with moderate sensitivity and poor specificity. Overreliance on CAD should be avoided and may actually degrade overall reader performance if used incorrectly. CAD should never be used to discount a finding deemed by the radiologist to be suspicious.